An X-ray of 2D Ear Biometrics

snakesailboatSecurity

Feb 23, 2014 (3 years and 7 months ago)

308 views

An X
-
ray of 2D Ear Biometric
s


Department of
Computer Science

Universiti Teknologi Malaysia,

81310 Skudai, Johor, Malaysia



ABSTRACT
:

Nowadays, it is obvious that computer have eye and can see beyond human imagination since the birth of biometric.
Basically, it was recently that human ear started gaining ground for identity authentication in biometrics because of its
rumple stable structure which makes a notable icon for feature extraction and this has spur computer vision authors on
the quest to de
veloping different algorithms for this purpose. Within this short space of time many questions has risen
due to the quality of the ear image used for recognition. They include environmental lighting condition, occlusion (due
to ear
-
rings, eye glasses, hair
), and pose angle and in trying to deal with the questionable issues many algorithms have
evolved. They follow some fundamental work
-
through approach; ear pre
-
processing/normalization, feature vector
extraction, classification and recognition. These quadru
ple are chained by a high bonding capability because none of
these stages can be compromised if a good ear recognition system is to be operational. This paper X
-
ray the state of 2D
ear normalization, detection and features selection method used for recogni
tion as documented by different research
works.

Keywords

-

Biometric, normalization, feature extraction, ear detection, ear
recognition.


I
ntroduction
:


It is pertinent to know that computer has eyes and can see, thanks to birth of Biometric
technology. Biometrics is a science
-
based authentication system which uses physical,
behavioural and chemical trait of human being for identification. These characteri
stics
include ear, fingerprint, iris, palmprint, signature, voice, odour, and DNA. It evolved in the
19th century due to proliferation of impersonation case witnessed from 16th century till that
time and this necessitated the use of fingerprint as biometr
ic character
[
1
]
. And since then,
the use of biometrics for various security and criminal check has risen exponentially due to
its success in authentica
tion process. Furthermore, this has saw major government facilities,
building and airport terminals equipped with different biometric scheme to combat crimes
like terrorism, information theft and electioneering process. The quest to conquer
fraudulent prac
tices several characteristics listed above was implored and ascertain viable
for recognition.



Figure 1: The Ear Anatomy


Human ear is a special organ located at both sides of the face which is used for hearing.
The ear is made up different parts; outer (Pinna), middle and the inner ear. For the purpose
of building a recognition system the Pinna, which the visible part of hum
an ear is used for
access authentication process.
[
2
]

viewed the ear as a small hearing organ with a multipart
which also aids human balance. This multipart make up the huge gullies which serve as a
pattern for recognition process and the parts includes the Helix, Fossa Triangularis, Crura
of Antihelix, Cym
ba Conchae, Anti Helix, Cacum Conchae, Tragus, Anti Tragus and
Lobule. Fig

1 depicts the outer ear (Pinna) with its different parts. Though the outer ear
(Pinna) shape is by nature made to assist in directing sound into the middle and inner ear,
biometric
basically uses it for identification because of its unique structure either in crime
scenario or access authentification purposes. Ear recognition system has become a major
focus in the biometrics world after its counterpart such as fingerprint, facial, an
d iris
recognition have been established to some great extent. This is because of its interesting
gully shape which remains constant over one’s life span unless in the case of accident.
Some of its merit includes, small geometric size, constant colour dist
ribution, and
unaffected by emotion or artefact such as creams
[
3
]
.


There are five stages involved in the Ear recognition scheme, they include: pre
-
processing (this includes Image nor
malization and edge detection), feature extraction,
classification and recognition stage. Each of these stages are as important as others, at such
none should be compromised.


2D E
ar

I
mage

A
cquisition
:

In order to confirm any new ideas in biometrics technology, the ear image need to be
acquired for experimentation and validation purposes. And since the position of the ear is at
the visible part of the body, there are no special cameras used for its acqui
sition, so any
camera can be used to capture the image for further processing. But these cameras should
be of a high resolution to ensure that good ear image is used for ear recognition process.
The image could be acquired with both monochrome and coloured

CCD cameras. Example
of some of the CCD cameras available in the market today is 365 by 240 pixeLINK
PLB959F
-
BF Monochrome CCD cameras and 360 by 360 pixel CP20 colour CCD camera
.
Furthermore, there are different standard Ear database used for purpose of
validation which
are also acquired by the cameras mentioned above. These databases are summarized in the
TABLE

1 and some of
the ear image is also shown in F
ig

2, this can be downloaded with
and without permission from the site
.

Table 1
:
Summary of Ear
Databases

Database
Name

Dimension Type

No of
Subjects

No of Ear
Images

Ear Image
description

USTB

University of Science and Technology Beijing

Database I

2D

60

180

Image captured
standard condition

Database II

2D

77

308

Image captured in
different
lighting
condition and angles

Database III

2D

79

1600

Image captured
with occlusion and
different angles


Database IV

3D

500

25500

Image taken with
the subject looking
eyelevel, upwards
and downwards.

UND

University of Notre Dame

Collection E

2D

114

464

Image taken with
different lighting
condition, pose angle
and days

Collection F

2D and 3D

302

942

Depth image

Collection G

2D and 3D

235

738

Profile/Depth
images

Collection J2

2D and 3D

415

1800

Profile/Depth
images

Collection
NDoff
-
2007

2D and 3D

396

7398

Different poses of
pitch and yaw

NCKU

National Cheng Kung University, Taiwan


3D face image

90

6660

Taken in different
angle

IIT Delhi

Indian Institute of Technology Delhi

Subset I

2D

190

801

Normal side image

Subset II

2D

89

801

Changes in

pose

WPUT

West Pomeranian University of Technology



2D

501

2071

Collected on
different
environmental
conditions.

SCface

Surveillance Cameras Face Database


3D face

130

4160

Collected on
different
environmental
conditions

XM2VTS

Extended Multi
-
modal Verification for Teleservices and Security


3D face

295

1180

Image taken with
different lighting
condition, pose angle
and days





Some sample ear image from WPUT database



USTB ear database sample



IIT delhi

ear database sample

Figure 3: Sample collection of 2D ear image from different database



E
ar

N
ormalization
:

All the existing algorithms of 2D ear recognition generally center its concept on either
frequency or spatial domain. Because of the variations in the images quality (due to the
difference in time, distance, angle, and lighting condition during acquisition
) in the
database used for validation, majority of the procedures starts by normalizing the image.
Normalization is a step taken to transform an image A to a standard form B so that all the
valuable information which can aid recognition retained and made p
rominent. In doing so,
variations are taken care of and a better performance in the detection/recognition scheme is
achieved with different database. The accuracy of extracted features from image depends
on the similarity quantities detected from such an

image and if the images are not similar
then the resultant feature vectors will be different. The ideology surrounding the birth of
normalization concept is because of the evolution of different ear databases used for
experimental validation of algorithm.

In the early days of ear biometrics, variation in the
image was not taken into consideration. So, the majority of the algorithm might not be
generalized because its result can be erroneous when used in another database. But today,
normalization to some ex
tent gives a consistent report on a database though with different
pose and yaw of the same subject. That is why
[
4
]

suggested that change in the image does
not make them to differ in quality but the disparity due to angle of capture and
environmental situation. Ear normalization by
[
5
]

used rotation principle by locating what
they called Max
-
line and calculating e
ar main angle, and then rotated the angle to 90
degrees. The length of the Max
-
line was used as the length of the ear for scale
normalization of the ear image. While in
[
6
]

involved only scale no
rmalization by resizing
the image from 300 by 400 to 100 by 150 after which they applied global thresholding for
compression purposes. In the same vein
[
7
]
, in trying to design an improve
d LLE,
converted colour image to grayscale and normalized the image to 45 by 24 pixels.


Meanwhile
[
8
]

observed that video frames always contain low quality image due to blur,
interlacing and compression. And they submitted that
, capturing of false features causes
erroneousness system, so they used force field transformation to smoothen the ear image.
Median filter was applied in
[
9
]

and linear interpolation to normalize the image
to 80 by
150 pixels. Statistical mean and variance in
[
10
]

was involved to normalize the ear image
quality for effective feature dominance. In 2004
[
11
]

was able to make his feature
extraction method invariant to both scale and rotation used the centroid of the ear image
as
a reference point to all other images to normalize the image, but the process on how this
centroid was found was not highlighted. Likewise
[
12
]

who normalized their image by
rescale the image to 90 by 60 and subsequently 30 by 30 for their sub images used for
feature extraction did not detailed their normalization method. I
n
[
13
]
, the image was
resized to 160 by 120 after cropping the image from side of the head and applied contrast
limited adaptive histogram. The normalization process by
[
6
]

involves resizing the image of
400 by 300 pixels to 100 by 150 pixels respectively and the application of global threshold
algorithm for the contrast. All the reports on normalization process above, points out its
importance in recognition process to ensu
re that image features from the same subject but
taken in difference scenario are consistent.


E
ar

D
etection
:

To some ear detection system is not a very major concern in 2D ear recognition since
there is compliance from the subject during the image capturi
ng but for real
-
time ear
recognition, it is adequately as it deals with moving frames and distance images. Generally,
object detection concepts makes use patterns and shapes which are created by the objects
spatial or frequency intensity. These basic ideas

is where 2D ear detected algorithm adopted
its control flow and this is mostly applied in a real time system and distance object where
system’s operator does not have control over his subject. To this end, there is no much
discussion on the ear detection

but some have presented a more qualitative report on this
issue. In 2012
[
14
]

proposed a connected component ear detection scheme which is
invariant to scale shape and rotation respectively. This was achieved by involving skin
segmentation approach to isolate the hairs that could be a problem to the system and with

edge computation and approximation after applying canny edge detection. Having the
concept that ear edge are convex in nature observes then and prune an
y noisy one and

Fig 4
shows the processes. Although the method reported an average accuracy rate of 97.
86%
and 96% with IITK and UND ear dataset. The problem of this algorithm is that it cannot
detect ear which is partially covered by either hair of scarf and poor quality image. A
curved shaped masking operator was proposed in
[
15
]

with four directions because of the
orientation of the ear shape for detection after some preprocessing activities and they went
further to validity it with Adaboost polling technique. The result of the system was
encouraging but under partial occlusion o
f hair and ear ring will result to a poor result.


More so,
[
16
]

were able to use light rays based on image ray transform to which assist in
extracting tubular features in highlighting the helix, and on introducing Gaussian
smoothing filter to reduce noise t be able to detect

the ear .They achieved this by taking
advantage of the helix elliptical shape, thereby reporting a 99.6% detection accuracy of
their algo
rithm on XM2VTS database. Fig
5 highlights the approach, from the application
of ray transform on the ear image to th
e detection f the ear image. It took
[
17
]

a cascaded
Adaboost classifier which has Haar features arranged
in it to detect the ear image within a
shortest possible time with an accuracy of 95% of a combination of different databases
(USTB, UMIST, UND, and WVHTF). Application of YCbCr skin
-
colour filtering in
[
18
]

with multi
-
template matching was used to accurately detect the dynamic and static ear with
~96% for the later, but they failed in the presence of poor quality image. Yet again, the
thrall of Adaboost and Haar feature took
[
19
]

into detecting the ear taking advantage of
concave
-
convex contour structure of the organ and detection was in two fold; off
-
line and
on
-
line respectively.

With the use of USTB, CAS
-
PEAL and CMU PIE databases for
validation, the report showed a promising result.



Figure 4: Ear Detection approach by 2012 Prakash and Gupa









Figure 5: Ray transform concept for ear detection by Cumming A. H. 2010


Meanwhile
[
20
]

used hierarchical clustering in locating the ear at the side of the face
by grouping the length of edges according clusters and those with long edge are assumed to
belong to the ear edge. Their result showed 94.6% accuracy on IITK database. In the same
ye
ar
[
21
]

took another voyage in automatic ear detection based on connected component
analysis constructed by usi
ng edge map of the ear with a high accuracy rate when used in
ear recognition process on IITK database. While
[
22
]

initiated a jet space similarity scheme
in detecting some major part of the ear which they acclaimed is the ear region as detected
with a better result as reported. It is observed that chain code aided the detection outer
curve of the helix in
[
23
]
, by detecting the edges before finding and reconstructing the helix
which is the outer border of the ear to be able to localize the ear. The experimental result
showed an accurate localization rate of 93.34% on IITK ear
database, but the method failed
in the presence of occlusion due hair. While
[
24
]

exploited the
average histogram of shape
index which aid them in their ear detection system with 91.5% accuracy rate. Also
[
25
]

applied elliptical Hough transform (HT) in feature extraction process to detect ear image.
Their experimental result showed a good performanc
e on XM2VTS face profile database
and UND ear database but could fail in the hair occlusion.

E
ar

R
ecognition
:

The contributions of many author is discussed in this section, highlighting the base
concept of the system and their likely drawback if subjected to a real life scenario and
basically only the 2D ear image is been discussed. Though
[
26
]

subdivided the recognition
into holistic, local, hybrid and statistical approach but we divided this process into two
basic g
roups; geometric and statistical based method. While the former makes use of the
structural shape of the ear, the latter involves a holistic approach. These two methods are
product of the types of feature extraction technique used in the recognition system
.

Geometric Based
:


In 1999
[
27
]

located six different points on the ear edge image and applied a heuristic
means to extract the feature
vectors at this points, and capture morphological vector which
they both feed into a compressed network which were fed into a weighting Bayesian
classifier for recognition process with 93% recognition accuracy rate. While
[
28
]

used
voronoi diagram which tak
es zero background pixel images and assigns them to nearest
seeds in a geometric distance of the contour for their recognition process. The error
correcting Graph match method was employed as similarity measure and their result was
promising, though with s
ome failure due to occlusion because the thermogram image
would hardly preserve the contour pattern. Geometric feature extraction scheme was
proposed in
[
29
]

which constructs circles with common center on the ROI of ear image and
extracted their radiuses

and the radius made by pixels at point of intersection with the
circle. He went further to extract the ends points of the contours and points where two
contours crossed themselves. Employing rule of thumb, their scheme was successful but
might fail in the

presence of occlusion due to earring and hair respectively. In the same vein
[
30
]

introduces a long axis based shape and structural feature extraction method
(LABSSFE) by drawing a line between the two longest point on the ear outer contour and
dividing the ear into two (Upper and Lower) lobes by locating the midpoint of the ini
tial
lo
ngest line drawn and Fig 6

s
ummarise the illustration. Then least square curve
approximation was employed to extract the features on the two lobes. Their second feature
extraction method localized the point of intersection between the two lines and the ear

contours. The result shows a success rate of 85% recognition accuracy under BP network
classifier, but might fail in a poor ear image quality.


Figure 6: The Outer ear contour with long axis by
Zhich
un Mu, Li Yuan et al. 2004



In recognition the human ear
[
31
]

proposed a model based ear identification system
which used morphological transform with cell automata to detecting the ear, before
iteratively determining the

contour with greater force of attraction. Determining the angles
made by some points on the contour feature vector were built which was subsequently
applied on a distanc
es measure for recognition. Fig

7 shows some of their processing
pedigree and thought
the system was a success, the ear
-
rings and hair would hamper its
efficiency. General Hough transform was used in
[
32
]

to extract four geometric features
points on the ear image and on applying a distance measure, their scheme showed a good
result thought the system might not perform well under occlusion. In 2007
[
33
]

came up
with another geometric method for ear recognition by developing an Improved Active
Shape Model which made use of the detected landmark

for normalization purpose to
iteratively search and locate several points on the edge of ear image.

Then using Full
-
spaced Linear D
iscriminant Analysis to recognize the ear image with 90% accuracy rate, it
is expected that the system might not perform wel
l under hair and earring occlusion.
Furthermore,
[
34
]

con
structed two reference points, the first they called max line and the
point of intersection of an orthogonal from the line to the outer ear contour was recorded as
a first feature vector. Another line was drawn from the top max line to meet the contour
dow
nward and same orthogonal line drawn as above; the angles of these points to the
contour are equally extracted.


Fig
ure

7: The feature extraction method used by Erno Jeg,Laszlo 2006


The F
ig 8

shows who they were able to effectively extracted the
feature vectors. On
applying rule of thumb, their experimental result show an accuracy recognition rate of 85%
but might be computational inexpensive and the technique failed in poor ear images as
reported by the researchers.


Figure

8: The Edge image and

feature extraction method by Jawale 2011


Statistical

Based:

By the use of geometrical concept
[
35
]

extract the force field transform information
from ear image and applying bi
-
directional Hausdorff distance on a collection of side face
profile images, an efficient recognition system was ach
ieved. The prior knowledge of Non
-
Negative matrix Factorization prompted
[
36
]

to develop an improved Non
-
Negative matrix
Factorization with sparsene
ss constraint for ear recognition technique in 2006. These
increase the dimension of the basis vector while minimizing the redundant feature
information thereby presenting a high quality feature for ear recognition and the recognition
rate increases with t
he increase in the feature vector dimension on Nearest Neighbour
classifier. The difficulty in accommodating rumpling of the ear structure
[
10
]

assembled a
multi
-

matcher ear authent
ication system. On sub
-
windowing the ear image the noisy edge
created due to the rumpled ear was taken care of, and extracting feature using Gabor filter.
Having that this filters are always created at high dimension space they applied Laplacian
EigenMaps
for dimensionality reduction and subsequently most Discriminant sub
-
window
was chosen using Sequential Forward Floating Selection. Based on the nearest neighbour
classifier the sub
-
windows were used to recognize the ear image and Fig. 8 is the highlights
o
f their processing activities towards achieving a recognition system. The experimental
result showed about 80% of rank 1 performance.


A new model proposed by
[
37
]

proposed a new model
-
based approach which makes us
e
of Hough Transform for the localization of the ear, applying SIFT with the assistance of
their iterative model called epoch duplicate keypoints are removed and more prominent
ones are then used in k
-
nearest neighbourhood class
ifier for ear recognition. F
ig
9 show
s

the sub
-
window and the Gabor filter effect on the image. The model presented a good
recognition rate on XM2VTS database used for validation purposes, occluded ear image
hard to recognition.




Figure

9: Gabor filter and convolution result and Sub
-
widow selected by SFFS of Nani and
Lumini 2007


The decomposition of the ear image to horizontal, vertical and diagonally by
[
38
]

using
wavelet transform was the bases for their recognition concept. The PCA was applied on
these three independent feature matrices produced for dimension reductio
n and recognition,
applying rule of thumb they were able to make 95.0 and 90.5% recognition accuracy on
USTB database II and USTB database III respectively. While a new ear recognition method
was proposed by
[
39
]
, uses PCA on the 2D wavelet which was applied on the image in
three different directi
ons to detect three independent

features. On evaluating these wavelet
coefficients based on recognition rate on USTB database, the horizont
al direction
outperformed the rest with 90.19% accuracy. A system
-
compounded structure classifier
approach was proposed in
[
40
]

for ear recognition. They
succeeded

in classifying the ear
into five groups based on the height to width

rate approach. Using PCA and ICA for further
feature vector extraction and recognition suggests that the latter performs better than PCA
on the three USTB ear database. The new local matching approach proposed by
[
41
]

made
use of Local Similarity Binary Pattern(LSBP), which makes use of both connectivity and
similarity information of the ear image for recognition. Before then, they used Cellula
r
Neural Network (CNN) for prepro
cessing purposes and on using nearest neigh
bour
classifier and Chi square the experi
mental result shows a good performance when
compared with similar well known approach. The
quest for reducing the computat
ional cost
of traini
n
g data during ear recognition necessitated
[
42
]

to

developed a wavelet transform
and discriminative common ve
c
tor(DCV) based ear recognition system. The Haar wavelet
transform decomposes the ear image into four sets components; high
-
frequency(x and xy
-
direction) and low frequency(x and xy
-
direction), redu
cing the dimension with DCV and
applying Euclidean distance for similarity measure. On testing the system and comparing
with PCA combination with LDA, they confirm the better accurate performance of their
system.


In strengthen the edge of ear image
[
43
]

was able to use force field transformation
before employing null space based kernel fisher discriminant analysis method to extract
feature vector that was used for recognition process. Using nearest neighbour distance
classifier the experimental result sh
owed an effective multi
-
pose ear recognition system on
USTB ear database. A SIFT point matching system was presented by
[
44
]

for 2D
registr
ation and recognition. This was achieved by taking advantage of the 2D nature of the
ear image used to create homography transform using SIFT and this system was said to
perform better when compared with PCA on XM2VTS database. Still in the search for
solu
tion on multi
-
pose ear recognition based system,
[
7
]

used the improved locally linear
embedding for this purpose since the locally linear embedding finds it difficult to select the
best neighbour in a higher di
mension sparse space. And USTB ear database certify the
effectiveness of the improved locally linear embedding against locally linear embedding on
Neural Network classifier with a promising recognition rate. While
[
45
]

proposed a fusion
of both local and global feature for their ear recognition technique. Kernel principal
component

analysis and independent component analysis were use for global and local
feature extraction respectively and linear SVM classifier was used. The canonical
correlation features was extracted based on correlation criterion function between two
groups of fe
atures, thereby establishing a Discriminant vectors that was used for ear
recognition. Their experimental result on USTB ear database displayed a better
performance.


The exploitation of the frequency boundary of the ear image by

[
46
]

using log
-
Gabor
filter thereby capturing its outer shape structure. Applying wavelet transform they were
able to sort for some of the occlusion problem and nearest neighbour algorithm were used
on Euclidean distanc
e for recognition purpose making their scheme show an effective result
on XM2VTS head profile image. The kernel independent components analysis provided the
required feature vectors for
[
47
]

in their ear recognition system before applying support
vector machine with Gaussian radial function in recognition process. The use of SVM with
Gaussian radial basis function their experimental result on Carrera
-
Perpinan
database
showed an encouraging
results
. While block segmentation technique for ear recognition
was conceived in
[
48
]

by dividing the ear image into sub
-
blocks and extracting feature
vectors using Statistical, Moment, Fourier and Gabor

feature respectively. The use of
Nearest Neighbour classifier on each of these feature vectors showed 100% recognition rate
for Moment feature vector on USTB ear database. A SIFT descriptor based ear recognition
system was proposed in
[
49
]
, on trying to make use of the SIFT descriptor on e
ar image
which has similar colour distribution, they introduced global context to reduce the
mismatch when there are many local SIFT descriptor similarity and finally used projective
invariant principle to reduce the problem of camera projection on the ear

image. Fig 10
shows their descriptor matching capability. The nearest neighbour classifier and USTB ear
database were used for the experimental validation with a recognition rate of 96.10%
accuracy. In the same vein
[
50
]

tried to effectively used SIFT descriptor for ear recognitio
n
process by proposing a robust system with fused SIFT descriptor from colour segmented
form slice regions of the ear image. Considering that the ear image is a collection of
coherent regions with Gaussian distribution representation they applied Gaussian
mixture
model and they were able to extract SIFT features from the colour slices. Using
concatenation and Dempster
-
Shafer theory they fused the features separately. The system
reported a 98.25 % identification rate on IIT Kanpur ear database. A neural netw
ork
integrator model was designed in
[
6
]

for ear recognition. Wi
th the zeal to increase the
computational speed of the ear recognition system they conceive making use of 2D wavelet
with global thresholding method and integrated neural network model with Sugeno
measure with Inner
-
Takes
-
All. The system was said to have 9
1.85% recognition rate
accuracy on USTB ear database.


Fig
ure 10: The SIFT descriptor

and the matching prowess of HuiZang,ZhiMu and
ShuaiWang 2009 work.


Meanwhile
[
9
]

on trying to sort for the il
lumination and posture problem of the ear
recognition system proposed applying isometric mapping algorithm for feature extraction,
since they can map dataset in low dimensional space through distance mapping. Using
fisher classifier, USTB ear database sho
wed a 100% accuracy rate in an illuminated ear
image. HERO was developed in
[
51
]

to carter for environmental changes in ear image
during ear recognition process. Making use of Haar wavelet and Partitione
d Iterated
Function System, the fractal indexing phase was exploited by HERO and with peano key
theory a more concise recognition system with less occlusion problem achieved. Testing
the algorithm on UND database showed a robust system under occlusion. Gab
or filter and
General discriminant Analysis was used in
[
52
]

for human ear recognition. While the Gabor
filter was used for feature extraction process because of its selectivity ability, the General
discriminant Analysis classified the features for recognition process. The experimental
report showe
d that the combination of the two outperformed the individual usage of the
algorithm with 99.1% recognition rate. The local Gabor wavelet in
[
53
]

was used to extract
feature vector a
nd they employed K
-
nearest neighbour which is a lazy learning method for
classification and recognition. The system achieved recognition accuracy rate of 91.51% on
USTB database. A quaternionic quadrature filter and quaternionic code were developed in
[
54
]

for the purpose of ear recognition, taking advantage of both the spatial and frequency
domain information present on the ear image. The system was said to have performed well
in UND and IITD ear database respectively
. Dividing the ear image into six regions,
[
55
]

employ the Eigen space fea
ture extraction method and consequently feed these features into
a multilayer neural network. There experimental result on USTB database showed a
promising result.




Figure
11: The widowing effect for feature extraction by Yuan and Mu 2012


To build an ear recognition system
[
56
]

took advantage of Gabor feature, they
segmented and normalize the 2D ear image using improved Adaboost and Active Shape
model principle for the detection purpose and further employ Kernel Fisher Discriminant
Analysis to reduce the high dimensionality crea
ted by Gabor feature. The system validation
with USTB ear database showed a cogent and promising result. An automated ear
identification system achieved in
[
57
]

by using morphological operator and Fourier
descriptor to localize the ear image before applying even Gabor filter extracting local
orientation information feature. T
hey went further to extract the more feature vector using
phase encoding using complex Gabor filter and with Adaboost detector the experimental
result on a public database showed 95.93 and 96.27% on a 221 and 125 subjects
respectively. I
n other to reduce t
he dimensiona
l
i
ty of the search space
[
56
]

had successful
experi
mental result in using Neighbourhood Preserving Embedding (NPE) for extracting
features from individual windows after subdivi
ding the ear image into sub
-
windows in their
ear recognition system. These features were later fused together to accommodate for any
par
tially occluded ear image. The F
ig 11
depicts

the sub
-
window set which they created
with overlapping principle and thoug
h the NPE is very good feature extraction approach,
but there will be too many features to account for and the ability to select the most
prominent ones will be a problem of this system.


C
onclusion
:

We have critically analyzed and presented a literature of 2D ear normalization,
detection and feature extraction technique for recognition. Some of the 2D ear database
were summarized in a table for easy consumption while highlighting the need to normalize

the ear which was taken under several conditions ranging from occlusion, angle of pose,
and environmental respectively. With the path taken by researchers, ear detection approach
became inevitably necessary when there is a search for full automation and i
ronically ear
recognition is more dependent on the feature extraction concept on which we have based
our grouping the recognition process into geometric and statistical ear recognition
approach. It is believed that this will serve as a pivotal leverage for

intending ear
recognition researchers to ameliorate their knowledge and existing models.


R
eferences:

Hurley, D.J., Force Field Feature Extraction for Ear Biometrics, in Electronics and
Computer Science

2001, University of Southampton: S
outhampton. p. 136
, 2001

Irwin, J., Basic Anatomy and physiology of the Ear, in HEARING PHYSIOLOGY, John
Wiley and Sons, ltd: USA.

2006

Xiu
-
qing, W., X. Hong
-
yang, and W. Zhong
-
li. The Research of Ear Identification Based
On Improved Algorithm of Moment In
variant. in Information and Computing (ICIC), 2010
Third International Conference on. 2010.

Adini, Y., Y. Moses, and S. Ullman, Face recognition: the problem of compensating for
changes in illumination direction. Pattern Analysis and Machine Intelligence,
IEEE
Transactions on, 19(7): p. 721
-
732
, 1997

Yazdanpanah, A. and K. Faez, Normalizing Human Ear in Proportion to Size and Rotation,
in Emerging Intelligent Computing Technology and Applications, Springer Berlin
Heidelberg. p. 37
-
45
, 2009.

Gutierrez, L., P
. Melin, and M. Lopez. Modular neural network integrator for human
recognition from ear images. in Neural Networks (IJCNN), The 2010 International Joint
Conference on. 2010.

Zhaoxia, X. and M. Zhichun. Ear recognition using LLE and IDLLE algorithm. in Patt
ern
Recognition, 2008. ICPR 2008. 19th International Conference on. 2008.

Cadavid, S. and M. Abdel
-
Mottaleb. Human Identification based on 3D Ear Models. in
Biometrics: Theory, Applications, and Systems, 2007. BTAS 2007. First IEEE
International Conference

on. 2007.

Sujuan, L., N. Jiangchuan, and F. Jun. Research into 2D ear recognition based on isomap
algorithm. in Industrial and Information Systems (IIS), 2010 2nd International Conference
on. 2010.

[Nanni, L. and A. Lumini, A multi
-
matcher for ear authent
ication. Pattern Recognition
Letters, 28(16): p. 2219
-
2226
, 2007.

Choraś, M., Ear Biometrics Based on Geometrical Method of Feature Extraction, in
Articulated Motion and Deformable Objects, F. Perales and B. Draper, Editors. 2004,
Springer Berlin Heidelber
g. p. 51
-
61.

[Yuan, L. and Z.c. Mu, Ear recognition based on local information fusion. Pattern
Recognition Letters, 33(2): p. 182
-
190
, 2012.

Tariq, A., M.A. Anjum, and M.U. Akram. Personal identification using computerized
human ear recognition system. in
Computer Science and Network Technology (ICCSNT),
2011 International Conference on. 2011.

Prakash, S. and P. Gupta, An efficient ear localization technique. Image and Vision
Computing, 30(1): p. 38
-
50
, 2012.

Hao
-
Chung, S., et al. Ear Detection Based on Arc
-
Masking Extraction and AdaBoost
Polling Verification. in Intelligent Information Hiding and Multimedia Signal Processing,
2009. IIH
-
MSP '09. Fifth International Conference on. 2009.

Cummings, A.H., M.S. Nixon, and J.N. Carter. A novel ray analogy for enro
lment of ear
biometrics. in Biometrics: Theory Applications and Systems (BTAS), 2010 Fourth IEEE
International Conference on. 2010.

Abaza, A., C. Hebert, and M.A.F. Harrison. Fast learning ear detection for real
-
time
surveillance. in Biometrics: Theory App
lications and Systems (BTAS), 2010 Fourth IEEE
International Conference on. 2010.

Heng, L. and L. Dekai. Improving adaboost ear detection with skin
-
color model and multi
-
template matching. in Computer Science and Information Technology (ICCSIT), 2010 3rd
I
EEE International Conference on. 2010.

Li, Y. and Z. Feng. Ear detection based on improved AdaBoost algorithm. in Machine
Learning and Cybernetics, 2009 International Conference on. 2009.

Prakash, S., U. Jayaraman, and P. Gupta, Ear localizatio
n using hier
archical clustering,

p.
730620
-
730620
, 2009

Prakash, S., U. Jayaraman, and P. Gupta. Connected component based technique for
automatic ear detection. in Image Processing (ICIP), 2009 16th IEEE International
Conference on. 2009.

Daishi, W., et al. Ear biome
trics using jet space similarity. in Electrical and Computer
Engineering, 2008. CCECE 2008. Canadian Conference on. 2008.

Ansari, S. and P. Gupta. Localization of Ear Using Outer Helix Curve of the Ear. in
Computing: Theory and Applications, 2007. ICCTA '0
7. International Conference on.
2007.

Hui, C. and B. Bhanu. Human ear detection from side face range images. in Pattern
Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference on. 2004.

Arbab
-
Zavar, B. and M. Nixon, On Shape
-
Mediated

Enrolment in Ear Biometrics

Advances in Visual Computing, Springer Berlin / Heidelberg. p. 549
-
558
, 2007.

A, P. and B. C, Ear biometrics: a survey of detection, feature extraction and recognition
methods. The Institution of Engineering and Technology, 1(2
): p. 114
-
129
, 2012.

Moreno, B., A. Sanchez, and J.F. Velez. On the use of outer ear images for personal
identification in security applications. in Security Technology, 1999. Proceedings. IEEE
33rd Annual 1999 International Carnahan Conference on. 1999.

B
urge, M. and W. Burger. Ear biometrics in computer vision. in Pattern Recognition, 2000.
Proceedings. 15th International Conference on. 2000.

Choras, M., Ear Biometrics Based on Geometrical Method of Feature Extraction. springer,
p. 51
-
61
, 2004.

Zhichun Mu
, et al., shape and structure feature based ear recognition. springer, p. 663
-
670
,
2004.

Jeges, E. and L. Mate. Model
-
Based Human Ear Identification. in Automation Congress,
2006. WAC '06. World. 2006.

Mahbubur Rahman, et al., Personal Identification Using

Ear Biometrics. International
Journal of the computer, the internet and Management, 15(2): p. 1
-
8
, 2007.

Li, Y. and M. Zhi
-
chun. Ear Recognition based on 2D Images. in Biometrics: Theory,
Applications, and Systems, 2007. BTAS 2007. First IEEE
International Conference on.
2007.

Jawale, J.B. and A.S. Bhalchandra. Ear based attendance monitoring system. in Emerging
Trends in Electrical and Computer Technology (ICETECT), 2011 International Conference
on. 2011.

Abdel
-
Mottaleb, M. and J. Zhou, Human
Ear Reco
gnition from Face Profile Image
Advances in Biometrics, Springer Berlin / Heidelberg. p. 786
-
792
, 2005.

Li, Y., et al. Ear Recognition using Improved Non
-
Negative Matrix Factorization. in
Pattern Recognition, 2006. ICPR 2006. 18th International Con
ference on. 2006.

Arbab
-
Zavar, B., M.S. Nixon, and D.J. Hurley. On Model
-
Based Analysis of Ear
Biometrics. in Biometrics: Theory, Applications, and Systems, 2007. BTAS 2007. First
IEEE International Conference on. 2007.

Nosrati, M.S., K. Faez, and F. Farad
ji. Using 2D wavelet and principal component analysis
for personal identification based On 2D ear structure. in Intelligent and Advanced Systems,
2007. ICIAS 2007. International Conference on. 2007.

Attarchi, S., M. Nosrati, and K. Faez, A New Ear Recognit
ion Approach for Personal
Identification, in Advanced Intelligent Computing Theories and Applications. With Aspects
of Theoretical and Methodological Issues, Springer Berlin Heidelberg. p. 356
-
363
, 2008
.

Haijun, Z. and M. Zhichun. Compound Structure
Classifier System for Ear Recognition. in
Automation and Logistics, 2008. ICAL 2008. IEEE International Conference on. 2008.

Yimo, G. and X. Zhengguang. Ear recognition using a new local matching approach. in
Image Processing, 2008. ICIP 2008. 15th IEEE In
ternational Conference on. 2008.

Hai
-
long, Z., et al. Ear recognition based on wavelet transform and Discriminative
Common Vectors. in Intelligent System and Knowledge Engineering, 2008. ISKE 2008.
3rd International Conference on. 2008.

Jiyuan, D. and M. Z
hichun. Multi
-
Pose Ear Recognition Based on Force Field
Transformation. in Intelligent Information Technology Application, 2008. IITA '08. Second
International Symposium on. 2008.

Bustard, J.D. and M.S. Nixon. Robust 2D Ear Registration and Recognition Bas
ed on SIFT
Point Matching. in Biometrics: Theory, Applications and Systems, 2008. BTAS 2008. 2nd
IEEE International Conference on. 2008.

Hai
-
Jun, Z. and M. Zhi
-
Chun. Ear recognition method based on fusion features of global
and local features. in Wavelet A
nalysis and Pattern Recognition, 2008. ICWAPR '08.
International Conference on. 2008.

Arbab
-
Zavar, B. and M.S. Nixon. Robust log
-
Gabor filter for ear biometrics. in Pattern
Recognition, 2008. ICPR 2008. 19th International Conference on. 2008.

Wu, H.
-
l., et

al. Ear Identification Based on KICA and SVM. in Intelligent Systems, 2009.
GCIS '09. WRI Global Congress on. 2009.

Wang, X. and W. Yuan. Human ear recognition based on block segmentation. in Cyber
-
Enabled Distributed Computing and Knowledge Discovery, 20
09. CyberC '09.
International Conference on. 2009.

Hui, Z., et al. Ear Recognition Based on the SIFT Descriptor with Global Context and the
Projective Invariants. in Image and Graphics, 2009. ICIG '09. Fifth International
Conference on. 2009.

Kisku, D.R.,
et al. An Efficient Ear Identification System. in Future Information
Technology (FutureTech), 2010 5th International Conference on. 2010.

De Marsico, M., N. Michele, and D. Riccio. HERO: Human Ear Recognition against
Occlusions. in Computer Vision and Patt
ern Recognition Workshops (CVPRW), 2010
IEEE Computer Society Conference on. 2010.

Xiaoyun, W. and Y. Weiqi. Gabor wavelets and General Discriminant analysis for ear
recognition. in Intelligent Control and Automation (WCICA), 2010 8th World Congress on.
20
10.

Tahmasebi, A., H. Pourghassem, and H. Mahdavi
-
Nasab. An Ear Identification System
Using Local
-
Gabor Features and KNN Classifier. in Machine Vision and Image Processing
(MVIP), 2011 7th Iranian. 2011.

Chan, T.
-
S. and A. Kumar, Reliable ear identification

using 2
-
D quadrature filters. Pattern
Recognition Letters, 2011.

Sholihah;, N.m., M.S.M. Rahim;, and D. Moamad, Ear recognition using hybrid features
and multilayer neural networks. International Journal of Computer Vision and Applications,
1(1): p. 24
-
30
,

2011.

Yuan, L., Z. Mu;, and H. Zeng, Ear Recognition based on Gabor Features ScienceDirect,
13: p. 3530
-
3537
, 2011.

Kumar, A. and C. Wu, Automated human identification using ear imaging. Pattern
Recognition, 45(3): p. 956
-
968
, 2012.